Decision LM
A Decision Language Model (Decision LM) is an AI model built to decide and predict rather than write. Instead of generating output text token by token, it reads an input, runs one pass through the model (equal to generating one output token), and returns a set of probabilities for that next token: which are then processed into a yes or no label, a set of confidence scores across classes, or a number in a range.
This way, we keep the language understanding of an LLM, but by stripping its ability to recurse (predict token after token till stop condition) and only emitting a single output over a limited vocabulary (your set of options), we make it exponentially faster (~50-500 ms E2E latency), cheaper (under $0.10/M tokens), and predictable enough to sit in pipeline relying on well-formatted outputs.
You may also see them called System One models (popularized by Daniel Kahneman's Thinking Fast and Slow & Jev), LLM classifiers, or Decisions API (OpenAI). They all mean the same thing.
What Decision LMs are used for
A Decision LM handles three kinds of task.
| Task | Output | Example |
|---|---|---|
| Binary classification (Noul/Predicate) | One probability that a statement is true | Does this response answer the user's question? |
| Multi-class classification (Choice) | A probability per class, summing to 1 | Which model should handle this query? |
| Regression (Score) | A single number on a scale you define | How urgent is this ticket, from 0 to 1? |
In practice these show up in a few recurring places:
- LLM judges: score outputs against criteria such as helpfulness, safety, or factuality.
- Content classification: moderation, intent detection, sentiment, and ticket triage.
- Model / Intent routing: send each query to the right-sized model or agent.
- Guardrails and verifiers: check an agent's input or output before the next step runs.
Not sure where a Decision LM fits in your own agent? Use the Discover your own prompt to have Claude Code find the spots for you.
Why use one
If the answer to a question is a choice among known options, a frontier model is overkill. A Decision LM gives you:
- Low latency. One forward pass, typically 50-500ms, with no generation loop.
- Low cost. Orders of magnitude cheaper per request than prompting a frontier model for the same call (~$0.04-$0.10/M tokens --> 2% of Opus)
- Confidence scores. Every decision comes with probabilities you can threshold, log, and route on.
For example, an Emissary judge returns in about 56 ms at $0.29 per 10k requests, where a frontier model takes 1.6–2.6 seconds and $10–15 per 10k. See the LLM judge and router quickstarts for the side-by-side numbers.
The tradeoffs, including calibration and input brittleness, are covered in Advantages and challenges. To see how Emissary stacks up against other Decision LMs on cost, latency, security, and calibration, read Comparing Decision LMs.
Where to go next
- Want to see one work? Start with the LLM judge quickstart or the router quickstart.
- Want to know how they work? Read How it works.
- Choosing a provider? See Comparing Decision LMs.
- Want to see where they fit? Browse the use cases.
- Zero-shot not accurate enough? See When to train your own.
- Want to call the API directly? See Experiments for binary, multi-class, and regression snippets.